Sociodemographic and geographic variation in mortality attributable to air pollution in the United States

Abstract

There are large differences in premature mortality in the USA by racial/ethnic, education, rurality, and social vulnerability index groups. Using existing concentration-response functions, particulate matter (PM_2.5) air pollution, population estimates at the tract level, and county-level mortality data, we estimated the degree to which these mortality discrepancies can be attributed to differences in exposure and susceptibility to PM_2.5. We show that differences in mortality attributable to PM_2.5 were consistently more pronounced between racial/ethnic groups than by education, rurality, or social vulnerability index, with the Black American population having by far the highest proportion of deaths attributable to PM_2.5 in all years from 1990 to 2016. Over half of the difference in age-adjusted all-cause mortality between the Black American and non-Hispanic White population was attributable to PM_2.5 in the years 2000 to 2011.

Despite improvements in the overall life expectancy in the USA over the past decades, significant inequalities among different racial/ethnic and socioeconomic groups remain a public health challenge¹. For example, in 2019, the life expectancy at birth was 78.8 years for the non-Hispanic (NH) White population but only 74.8 years for Black Americans².

Air pollution, particularly exposure to fine particulate matter (PM_2.5), is thought to be a major risk factor for premature death, both worldwide and in the USA³. Although most areas in the USA have seen declines in air pollution over the past decades, the pollution-related health burden remains substantial³. Improvements in air pollution exposure have been unequally distributed across population subgroups^4–6. For instance, Tessum et al.⁵ showed an overall decrease in PM_2.5 exposure in the USA between 2005 and 2015, but that such benefits were less pronounced among Black Americans, communities with lower educational attainment, and rural communities. Several federal environmental policies have been implemented since the 1970 Clean Air Act to address the fact that air pollution exposure differs across USA communities and acknowledge the environmental justice implications of exposure to air pollution. In particular, Executive Order 12898 (focusing on identifying and addressing the disproportionately high adverse human health effects from environmental exposures and on developing strategies for implementing environmental justice), Executive Order 14008 (focusing on climate change-related impacts and encouraging mitigations strategies that may also have co-benefits on greenhouse gas emissions and public health) and the National Ambient Air Quality Standard program state that air quality standards must be set at a level that protects the most vulnerable populations. More information on these policies is provided in the Supplementary Text.

In addition to being systematically more exposed to higher levels of air pollution, structurally disadvantaged communities are also thought to be more susceptible to adverse health effects from air pollution, which has sometimes been referred to as an environmental justice “double jeopardy”⁷. For example, the concentration-response function (CRF) linking PM_2.5 exposure to mortality appears to be more pronounced among Black Americans⁸. Such differential susceptibility is thought to be due to differential distributions in pre-existing comorbidities, lack of access to health care, racialized occupational sorting into jobs with more hazardous exposures, and other social and structural determinants of health⁷. Indeed, there is a rich literature documenting how structural racism⁹, related to historical policies such as redlining as well as contemporary inequalities in health care access, can explain such differential susceptibility to air pollutants¹⁰.

Using existing CRFs, PM_2.5 air pollution and population estimates at the census tract level, and mortality data at the county level, the aim of this study was to quantify the contribution of PM_2.5 to racial/ethnic, educational, geographic, and social vulnerability-related inequalities in mortality in the USA to inform appropriate policy interventions. Studies that estimate the number of deaths attributable to air pollution, and PM_2.5 in particular, usually rely on existing CRFs from the literature to infer such burden¹¹. These studies have used a single CRF for the whole pooled population, assuming no differences in susceptibility to PM_2.5, which may drastically underestimate the unequal health burden of air pollution across vulnerable communities. Given the evidence on differences in susceptibility to PM_2.5 by racial/ethnic groups in the USA ^9–14, we used the CRF from Di et al.¹⁵ for our primary analyses. The CRF from Di et al. is the only existing CRF that provides race-ethnicity-specific estimates for each of the main racial/ethnic categories in the US census and is based on a large sample of the population from all US counties¹⁵. Our analysis relies on the assumption that this CRF accurately depicts a causal effect of PM_2.5 exposure on mortality. To examine the degree to which the use of CRFs that ignore differences in the concentration-response association by racial/ethnic group underestimates disparities in PM_2.5-attributable mortality between racial/ethnic groups, we then compare our results to those obtained from two recent and widely used uniform CRFs for the USA population in secondary analyses ^11,16. In this study, we use the classification of racial/ethnic groups as a proxy for long-established and systemic consequences of political, historical, and economic structures, social constructs as well as environmental racism^10,12,17.

Results

Exposure to PM_2.5 at the national level:

In 2016, 0.8% of all census tracts and 0.9% of the overall population were exposed to an annual mean PM_2.5 concentration above 12µg/m³, which is the legally required threshold set by the current National Ambient Air Quality Standard¹⁸. In comparison, in 1990, 83.4% of census tracts and 85.9% of the population were exposed to PM_2.5 levels above 12µg/m³. The decline in PM_2.5 exposure (as well as all-cause mortality) overall, and separately by subpopulation over time is plotted in figures Fig. S 2 -Fig. S 7. The mean population-weighted PM_2.5 exposure levels, when averaged over the period from 2000 to 2016, were highest among Black Americans at 9.38 μg/m³, followed by Asians or Pacific Islanders at 9.21 μg/m³, Hispanics or Latinos at 9.13 μg/m³, Non-Hispanic Whites at 8.25 μg/m³ and American Indians or Alaska Natives at 7.46 μg/m³ (Fig. S 4). Between 2000 and 2016, the disparity in PM_2.5 exposure between Black Americans and Non-Hispanic Whites narrowed from 2.17 μg/m³ to 0.94 μg/m³.

Estimated PM_2.5-attributable mortality at the national level:

In the overall population, the estimated PM_2.5-attributable mortality declined from 79.2 (95% CI, 77.1 to 81.4) age-adjusted deaths per 100,000 in 1990 to 11.7 (95% CI, 11.4 to 12.0) in 2016. We observed this steep downward trend for all studied subpopulations, with absolute differences between these groups narrowing considerably over the study period (Fig. 1, Fig. S. 8). However, the lines representing each subpopulation in figure 1 did not intersect at any time between 1990 and 2016. Thus, the most affected groups in 2016 were also the most affected groups in 1990.

Fig. 1. Age-adjusted PM_2.5-attributable mortality rate by racial/ethnic group, education level, rurality level, and the social vulnerability index.

The first row shows the age-adjusted mortality rate that we estimate is attributable to PM_2.5. The second row shows the percentagae of all-cause mortality that we estimate is attributable to PM_2.5. The dashed lines depict 95% confidence intervals. 95% confidence intervals for rurality are too narrow to be visible.

Abbreviations: NH=Non-Hispanic, SVI=social vulnerability index.

From 1990 to 2016, Black Americans experienced an estimated 86% reduction in the age-adjusted mortality rate attributable to PM2.5, the largest decline observed among all racial/ethnic groups (Fig. 1). However, these larger improvements among Black Americans must be considered in relation to their higher starting point in the PM_2.5-attributable mortality rate relative to other racial/ethnic groups. In fact, although the absolute size of the difference in the PM_2.5-attributable age-adjusted mortality rate between Black Americans and other racial/ethnic groups decreased over the study period, the relative size of these differences remained similar over time (Fig. S 9). Examining the estimated percent of all-cause mortality that is attributable to PM_2.5, rather than the PM_2.5-attributable mortality rate, for each population group yielded similar trends (Fig. 1). Of note, the declines in PM_2.5 exposure over the study period have led to a decrease in the percent of all-cause mortality that is attributable to PM_2.5 among each of our population subgroups. Unlike the decline in the PM_2.5-attributable mortality rate over time, this finding was not self-evident because it indicates that for each population subgroup mortality from PM_2.5 decreased more rapidly than mortality from other causes.

More than half of the difference in all-cause mortality between Black Americans and the Non-Hispanic White population was attributable to PM_2.5 in each year from 2000 to 2011 (Fig. 2). With a decrease from 53.4% (95% CI, 51.2% to 55.9%) in 2000 to 49.9% (95% CI, 47.8% to 52.2%) in 2015, this proportion, however, has declined over time. The percent of the difference in all-cause mortality to Black Americans that can be explained by differences in exposure and susceptibility to PM_2.5 between racial/ethnic groups was lower for the Hispanic or Latino White, Asian or Pacific Islander, and American Indian or Alaska Native population than for the Non-Hispanic White population. Nonetheless, at a mean for the period 2000 to 2015 of 20.8% (95% CI, 20.7% to 20.9 %), 16.1% (95% CI, 16.1% to 16.2%) and 12.8% (95% CI, 12.7% to 12.8%) for the American Indian or Alaska Native, Hispanic or Latino White, Asian or Pacific Islander, and population, respectively, the percentages were still substantial.

Fig. 2. Extent to which the difference in the age-adjusted mortality rate between each racial/ethnic group and Black Americans can be attributed to PM_2.5.

Of the racial/ethnic groups considered in this study, Black Americans have both the highest all-cause mortality rate and highest PM_2.5-attributable mortality rate. The first column is the absolute size (in mortality per 100,000) of the difference in mortality attributable to PM_2.5 between each racial/ethnic group and Black Americans. The second column is the percent of the difference in all-cause mortality between each racial/ethnic group and Black Americans that can be attributed to PM_2.5 exposure.

Dark blue dots denote 2015 and light blue dots denote 2000. Red dots denote the unweighted mean from 2000 to 2015. The dot sizes are proportional to the age-adjusted all-cause mortality rate that is attributable to PM_2.5.

Abbreviations: NH=Non-Hispanic

We observed that disparities between racial/ethnic groups in PM2.5-attributable mortality exist at all levels of education, rurality, socioeconomic status, household characteristics, housing type and transportation, and social vulnerability index (Figures 3, S 10 – S 15). In fact, variation in PM2.5-attributable mortality by racial/ethnic group further increased when the analysis was restricted to the population with high educational attainment or those residing in non-metropolitan areas (Fig. S 16). The variation in PM2.5-attributable mortality by racial/ethnic group was similar for the different levels of socioeconomic status, social vulnerability index, and housing type and transportation (Fig. S 16). Black Americans and Hispanic or Latino Whites had the highest PM2.5-attributable mortality at all levels of these variables in the year 2016 (Figures 3, S 10 – S 14). Black Americans also had a higher PM2.5-attributable mortality than the respective average for every level of these variables in 2016 (Figures 3, Fig. S 10–Fig. S 14). Hispanic or Latino Whites living in socially vulnerable counties experienced double the PM2.5 attributable mortality than Hispanic or Latino Whites living in socially resilient counties in 2016 (Fig. S 10, Fig. S 17).

Fig. 3. Age-adjusted PM_2.5-attributable mortality rate and all-cause mortality rate for each racial/ethnic group stratified by educational attainment.

The dashed lines depict 95% confidence intervals. The US Census Bureau has not published data on educational attainment by age for the racial/ethnic group “American Indian or Alaska Native”³⁴. This racial/ethnic group has, thus, been omitted from this figure.

Abbreviations: NH=Non-Hispanic.

Those with low education and Black Americans had a higher PM2.5-attributable mortality than the respective average for every level of rurality, socioeconomic status, housing type and transportation, and social vulnerability index in 2016. Disparities in PM2.5-attributable mortality by educational attainment were more pronounced for high socioeconomic status than for middle and low socioeconomic status (Fig. S 18, Fig. S 19). Those with a high school diploma or a lower level of education experienced a higher PM2.5-attributable mortality than those with a higher educational attainment at all levels of rurality, social vulnerability index, socioeconomic status, household characteristics, minority status, and housing type and transportation (Fig. S 19 - Fig. S 24). High school graduates or lower living in large metropolitan areas experienced 36.9% (95% CI, 34.5% to 39.3%) higher PM_2.5 attributable mortality than those with the same level of education living in non metro areas in 2016 (Fig. S 25).

Disparities in PM_2.5-attributable mortality by education or rurality were not as pronounced as for racial/ethnic groups (Fig. 1). This finding was also corroborated when adapting a coefficient of variation approach (15), which found that estimated age-adjusted PM_2.5-attributable mortality varied more by racial/ethnic group than by education or rurality in all years of our study period (Fig. S 26). Nonetheless, differences in PM_2.5-attributable mortality by education and by rurality were still apparent. In the year 2016, those with a high school diploma or lower experienced 16.9 (95% CI, 16.4 to 17.4) PM_2.5-attributable age-adjusted deaths per 100,000 (Fig. 1). This rate is nearly double compared to those with some college education but no 4-year degree, who had a rate of 8.8 (95% CI, 8.5 to 9.0), and those with a 4-year college degree or higher, who had a rate of 7.7 (95% CI, 7.5 to 7.9) (Fig. 1). Similar patterns are observed in all-cause mortality rates (Fig. S 5). Between 1990 and 2016, there was a significant reduction in the absolute differences in PM_2.5-attributable age-adjusted mortality rates per 100,000 between large metro and non-metro areas, declining from 28.7 to 5.3. Despite this decrease in absolute disparity, the relative difference remained stable and even showed a slight increase (Fig. 1). Specifically, large metro areas had 1.48 times the PM_2.5-attributable mortality rate of non-metro areas in 1990, and this ratio increased to 1.57 times in 2016 (Fig. 1).

Variation in estimated PM_2.5-attributable mortality across states and counties:

We observed substantial variation across states both in the age-adjusted PM_2.5-attributable mortality rate (Fig. S 27) as well as the percent of age-adjusted all-cause mortality that can be attributed to PM_2.5 (Fig. 4). Fig. S 28 depicts the state-level positive association between PM_2.5 exposure and all-cause mortality. In all states with a PM_2.5-attributable mortality rate above zero, Black Americans had a higher percent of all-cause mortality that can be attributed to PM_2.5 (Fig. 4), as well as a higher PM_2.5-attributable mortality rate (Fig. S 27), in 2016 than the Non-Hispanic White population. Similarly, those with a high school diploma or lower education had a higher PM_2.5-attributable mortality in all states compared to groups with a higher level of education (Fig. S 28). In 31 of the 34 states that had both counties designated as “large metro” and “non metro”, “large metro” counties had a higher PM_2.5-attributable mortality compared to “non metro” counties (Fig. S 28).

Fig. 4. Percentage of the age-adjusted all-cause mortality rate that was attributable to PM_2.5 in the year 2016, by state and racial/ethnic group.

Abbreviations: NH=Non-Hispanic.

When examining variation across counties, we found that in virtually all (96.6%) counties for which data were available, Black Americans had a higher age-adjusted PM_2.5-attributable mortality rate than the Non-Hispanic White population (Fig. 5, Fig. S 29) for the mean value across the period 2000 to 2016. The comparison between the Black American and Hispanic or Latino White population was more varied, with a number of counties in the South-Eastern USA having a lower PM_2.5-attributable mortality rate for Black Americans than for Hispanic or Latino Whites. The lowest PM_2.5-attributable mortality rates for all three racial/ethnic groups for which we had a sufficient sample size at the county level (Black American, Hispanic or Latino White, and Non-Hispanic White) tended to be in counties in the Mountain West. When taking into account spatial autocorrelation, we identified multiple county clusters with significantly higher or lower differences in the PM_2.5-attributable mortality rate between racial/ethnic groups than in surrounding areas (Fig. S 30).

Fig. 5. Differences in the age-adjusted PM_2.5-attributable mortality rate between racial/ethnic groups at the county level for the period 2000 to 2016.

These figures show the (unweighted) mean value across the period 2000 to 2016.

The first row is the PM_2.5-attributable mortality per 100,000 for each racial/ethnic group.

The second row is the absolute difference in PM_2.5-attributable mortality per 100,000 between racial/ethnic groups.

Abbreviations: NH=Non-Hispanic.

All our analysis results, including differences in PM_2.5-attributable mortality by racial/ethnic, education, level of rurality, and social vulnerability-related groups, as well as trends over time, remained similar in relative terms when restricting the study population to those aged 65 years and older (Figures S 31 – S 45).

Discussion

Our study shows that improvements in air quality in the USA have decreased estimated PM_2.5-attributable mortality for all subpopulations that we examined. However, our analysis also highlights the remaining inequalities in PM_2.5-attributable mortality between different subpopulations. These inequalities were most pronounced between racial/ethnic groups. In fact, more than half of the difference in mortality between the Non-Hispanic White population and Black Americans was attributable to PM_2.5 in each year from 2000 to 2011.

Our results indicate the strong association of the race and ethnicity group with adverse environmental health outcomes – an association that is even stronger than for education, rurality or social vulnerability-related factors. This finding aligns with a growing body of evidence ^9,10,19 demonstrating that racial/ethnic categories are not simply proxies for socioeconomic differences but are also (imperfect) proxy measures for exposure to historical and contemporary discriminatory practices. Racism and the discrimination of racial/ethnic minorities emerged early in the founding of the USA and evolved into government-sponsored displacement, exclusion, and residential segregation^9,10,19. Residential segregation is a substantial consequence of structural racism that still negatively impacts the health of racial/ethnic minorities today through various mechanisms involving both a disproportionate exposure to air pollutants and higher susceptibility to these pollutants^{4,7,9,15,20–22}. Sources of air pollution emissions are often located in marginalized communities, as residents of these areas tend to have less economic opportunity, resources, and social capital, as well as limited political power, to influence the decision-making processes that determine where such sources of pollution are placed^9,10,20. In addition to its role in explaining differences in PM_2.5 exposure between racial/ethnic groups, structural racism is likely also a major driver of the unequal distribution of factors that contribute to the, on average, higher susceptibility to PM_2.5 exposure among many racial/ethnic minorities¹⁵. These factors include social determinants of health, such as exclusion from job and educational opportunities, inadequate access to healthcare, and maladaptive coping behaviors⁹.

An additional important finding of our analysis is that inequalities in PM_2.5-attributable mortality between racial/ethnic groups were obscured when ignoring differences in the estimated susceptibility to PM_2.5 across racial/ethnic groups. Using race-ethnicity-specific mortality rates and PM_2.5 exposure measurements, but assuming that the mortality effects of a given unit of PM_2.5 are equal across all racial/ethnic groups, will greatly underestimate differences in PM_2.5-attributable mortality. Similarly, using CRFs that do not take into account differences in susceptibility to PM_2.5 between racial/ethnic groups will generally underestimate the benefit of reductions in PM_2.5 on health outcomes among structurally disadvantaged groups. We, thus, advocate for the use of race-ethnicity-specific CRFs in future health impact studies to ensure air quality policies protect subpopulations most at risk.

We note some caveats of our study. A first set of limitations relates to the CRFs that we used in our analysis. First, although the race-ethnicity-specific CRF by Di et al.¹⁵ used in this analysis was developed in the age group 65+ years, our manuscript presents results for the age group 25+ years. To investigate whether our results would differ substantially if we conducted our study in this older age group only, we have implemented all analyses shown in this manuscript when restricting the study population to the age group 65+. The results (shown in the Supplementary Materials) demonstrate that all relative differences and patterns over time by racial/ethnic, education, and geography group remain similar. Second, we use a uniform CRF to derive the mortality response for all subpopulations not related to racial/ethnic group. There is evidence of differences in susceptibility to PM_2.5 for the different subpopulations. For example, the composition of air pollution in urban areas has more adverse health effects than in rural areas²³. We may, thus, underestimate the differences in PM_2.5-attributable mortality between subpopulations not related to racial/ethnic group. Third, because the true underlying PM_2.5-attributable mortality rate is not practically measurable, we cannot estimate the validity of any given CRF. We therefore present results using multiple CRFs, selecting Di et al.’s model as our primary CRF due to its unique provision of race-ethnicity-specific estimates and its foundation in a nationwide sample^11,16,24.

A second set of limitations relates to our PM_2.5 exposure measurements. We relied on assigning PM_2.5 exposure based on the current location of residence. This could be problematic for two reasons. First, if people live and work at different locations and pollution levels are systematically different between these locations, our attribution estimates may be biased. However, available evidence for urban areas finds the discrepancies in work and residential PM_2.5 levels are small in absolute values (<0.1µg/m³)²⁵, suggesting that assigning exposures by residence location may be a reasonable approximation of total exposure. A related concern arises if mortality is driven mainly by long-term cumulative exposure, populations change residence locations during the study period, and there exists a systematic correlation between exposure at previous residence location and exposure at current residence location. However, existing literature finds that both immediate and long-term exposure matter for health outcomes, and again does not find consistent support for patterns of geographic mobility being associated with pollution levels in the USA²⁶. Another potential limitation associated with exposure assignment was due to pollution and population data being available at a geographically more granular level (0.01° by 0.01° and the census tract level, respectively) than our mortality data (available at the county level). As the geographic level at which we conduct our analysis increases, the risk of ecological bias increases. However, it is worth noting that the ‘smoothing’ effect of using county-level data is more likely to result in underestimation rather than overestimation of disparities between subpopulations. This is because the averaging process could dilute more extreme values that contribute to these disparities, making them appear less pronounced than they actually are. We, however, did not observe a substantial association between the proportion of all-cause mortality attributable to PM_2.5 with within-county variation in PM_2.5 (S32, Panels C), implying that aggregation of PM_2.5 exposure from the census tract to the county level is unlikely to be an important source of bias in our analysis. Another limitation is that our final confidence intervals for PM2.5-attributable mortality do not incorporate uncertainty arising from the PM2.5-exposure estimates, because the PM2.5-exposure estimates on a 0.01° by 0.01° grid from Meng et al. only provide mean estimates (without uncertainty estimates). Another limitation is that Di et al.’s CRF is derived from zip-code level data, while our study is based on coarser county-level data. Hence there is a difference in exposure measurement errors between those geographic levels. This could affect the validity of the point estimates and confidence interval of the CRF by Di et al. estimates when applied to our study. For the first point, a previous study²⁷ suggests that the bias for the point estimate introduced by the exposure measurement error is generally small.

A third set of limitations of our study relates to our measurement of socioeconomic variables. First, we had to rely on racial/ethnic group as a proxy measure for exposure to racism. Richer data on exposure to racism would be a key asset for future research in this area²⁸. Second, there is evidence that educational attainment recorded on death certificates tends to overestimate the decedent’s level of education²⁹. Lastly, because information on the social vulnerability-related factors was not available on the death certificates, we assigned the social vulnerability-related factors to all individual deaths within a county. It is important to note that contextual factors such as the social vulnerability index we used may not capture the individual-level socioeconomic status. Yet, contextual socioeconomic vulnerability may be a more relevant indicator to inform/guide targeted policies across the US counties to reduce air pollution related health disparities³⁰. That being said, it would be interesting to explore multi-level socioeconomic disparities in relation to deaths attributable to air pollution in future work as such data becomes available.

Methods

We harmonised mortality counts, population counts and PM_2.5 concentration estimates to estimate PM_2.5-attributable mortality across various sociodemographic population subgroups in the United States between 1990 and 2016. The population subgroups used in our analyses were racial/ethnic groups, educational attainment groups, rurality levels, socioeconomic status, household characteristics, racial and ethnic minority status, housing type and transportation, and social vulnerability index. We derived population counts at the census tract level from the U.S. Census Bureau, using linear interpolation to account for missing data in intercensal years. We acquired restricted-use county-level mortality counts from the U.S. National Vital Statistics System. Both data sets were pre-disaggregated by racial/ethnic groups and educational attainment at their respective sources. Additionally, we mapped levels of rurality, socioeconomic status, household characteristics, racial and ethnic minority status, housing type and transportation levels, and the social vulnerability index to the population counts and mortality counts based on county level look-up tables from the National Center for Health Statistics and the Centers for Disease Control and Prevention.

We mapped PM_2.5 concentration levels from an established model³¹ and ground-based measurement data³² to census tracts. We assigned population-weighted mean PM_2.5 exposure estimates to each county. To estimate the mortality burden that is attributable to PM_2.5 at the county (and, subsequently, state and national) level, we combined our annual population-weighted mean PM_2.5 exposure estimate and mortality counts at the county level with a CRF. The main CRF used in this analysis was by Di et al.¹⁵. For state or national-level analyses, we aggregated county-level all-cause and PM2.5-attributable mortality counts to the respective state or national level. We converted these raw mortality counts into age-adjusted mortality rates.

A detailed description of the data sources and methods can be found in the supplement.

Funding Statement

PG is a Chan Zuckerberg Biohub investigator. DF is supported by the Gerhard C. Starck Foundation. MVK is supported by National Institutes of Health (NIH) grant R00DA051534. EB is supported by NIH grants R01AI127250 and R01HD104835. SHN is supported by the Robert Woods Johnson Foundation. TB is supported by NIH grant R01CA228147 and by the California Environmental Protection Agency’s Office of Environmental Health Hazard Assessment (#21-E0018).